For an ensemble-based history matching of a channelized reservoir, loss of geological plausibility is challenging because of pixel-based manipulation of channel shape and connectivity despite sufficient conditioning to dynamic observations. Regarding the loss as artificial noise, this study designs a serial denoising autoencoder (SDAE) composed of two neural network filters, utilizes this machine learning algorithm for relieving noise effects in the process of ensemble smoother with multiple data assimilation (ES-MDA), and improves the overall history matching performance. As a training dataset of the SDAE, the static reservoir models are realized based on multipoint geostatistics and contaminated with two types of noise: salt and pepper noise and Gaussian noise. The SDAE learns how to eliminate the noise and restore the clean reservoir models. It does this through encoding and decoding processes using the noise realizations as inputs and the original realizations as outputs of the SDAE. The trained SDAE is embedded in the ES-MDA. The posterior reservoir models updated using Kalman gain are imported to the SDAE which then exports the purified prior models of the next assimilation. In this manner, a clear contrast among rock facies parameters during multiple data assimilations is maintained. A case study at a gas reservoir indicates that ES-MDA coupled with the noise remover outperforms a conventional ES-MDA. Improvement in the history matching performance resulting from denoising is also observed for ES-MDA algorithms combined with dimension reduction approaches such as discrete cosine transform, K-singular vector decomposition, and a stacked autoencoder. The results of this study imply that a well-trained SDAE has the potential to be a reliable auxiliary method for enhancing the performance of data assimilation algorithms if the computational cost required for machine learning is affordable.